Most airline personalisation conversations start in the wrong place: with the offer. A passenger sees a contextual bag upsell, the conversion rate ticks up, and someone declares the personalisation engine is working. It might be. But that bag upsell is the last step in a chain that begins much earlier, with far less glamorous work.
If your personalisation efforts feel like they're underperforming, have inconsistent uplift, low ancillary attach rates, or recommendations that miss the mark, the answer is rarely in the offer itself. It's almost always in the layers underneath it.
This is how the full stack fits together.
The five layers
Think of airline personalisation not as a single capability but as five interdependent layers, each one a precondition for the next:
- Segmentation: knowing who you're talking to
- Contextualised selling: deciding what to say to them
- Optimisation: continuously improving that decision
- Dynamic delivery: presenting it in the right format, in the right place
- Ancillary conversion: where it either pays out or doesn't
The architecture matters because the most common failure mode isn't a weak offer, it's a strong offer sitting on a broken foundation. If you skip layer 1, layers 2 through 5 are optimising noise.
Layer 1: Segmentation = the foundation everything else rests on
The industry default for passenger segmentation is largely demographic or booking-class based. Business traveller. Leisure family. Frequent flyer. These categories describe who someone looks like on paper. They don't predict what someone will do at checkout.
Behavioural segmentation changes the question entirely. Instead of asking "what type of traveller is this?", it asks "what has this passenger consistently chosen, and what does that predict about their next decision?" The difference sounds subtle but the impact isn't.
Airlines sitting on years of booking data have an enormous advantage here that most haven't fully used. The passenger who always books an aisle seat, always adds priority boarding, and always skips seat upgrades is telling you exactly what to offer them, and exactly what not to. When your segments are built on that signal rather than demographic approximations, the decisioning layer that follows gets materially more accurate.
The loyalty upside is equally significant. Segmentation that reflects actual behaviour creates triggers for relevant communication throughout the journey, not just at the point of booking. That's the difference between a loyalty programme that labels customers and one that genuinely drives repeat purchase.
Layer 2: Contextualised selling = the right offer, in the right moment
Once you know who someone is behaviourally, the next question is what to offer them, and precisely when. This is where contextualised selling does its work.
Context here means more than passenger identity. It means route characteristics, time to departure, what the passenger has already selected in the current session, and external factors like destination weather or local events. A lounge offer that converts well at booking converts very differently the morning of departure. The offer hasn't changed. The context has.
The critical insight from incentivising with contextualised ancillaries is one that revenue teams consistently underweight: relevance outperforms discount. A correctly contextualised full-price ancillary drives more conversion than a discounted product served to the wrong passenger at the wrong moment. Research on dynamic airline ancillary pricing backs this up: contextual models have shown a 36% increase in conversion rate compared to blanket discount approaches.
Suppression logic matters here as much as recommendation logic. Knowing what not to show, and avoiding the noise that erodes trust in your offers over time, is a core part of the decisioning layer, not an afterthought.
Layer 3: Optimisation = the system that gets smarter over time
The decisioning layer gives you your best current estimate of the right offer. Reinforcement learning is what stops it from staying an estimate.
Contextual bandits — the specific machine learning approach most applicable to airline personalisation — learn from live outcomes in production rather than only from historical training data. Every booking, every ancillary purchase, every skip updates the model. The system explores new combinations on a small slice of traffic to discover what works better, then exploits that learning across the majority. Over time, the gap between what the system thinks will convert and what actually converts closes continuously.
This is the layer most airlines are missing. A contextualised selling engine that isn't fed by a live optimisation loop is a static ruleset: good at launch, gradually stale. McKinsey's research consistently shows that companies with live personalisation optimisation generate up to 40% more revenue from those efforts than companies running fixed personalisation logic. That gap is almost entirely explained by Layer 3.
Layer 4 = Dynamic delivery: the offer has to land in the right format
You have the right offer for the right passenger, optimised in real time. Now it has to actually reach them in a way that converts.
This is where dynamic layout and dynamic presentation come in. They're doing different jobs, which is why conflating them is a costly mistake.
Dynamic presentation covers what is shown: which product, which image, which copy, which price format. Dynamic layout covers how the page is structured around it: which modules appear, in which order, at what size. Two passengers on the same flight, at the same fare level, can be served a structurally different page if their behavioural profiles are sufficiently distinct, and that difference in structure can move conversion by as much as the underlying offer itself.
The failure mode to avoid: implementing dynamic content (personalised images, copy variants) without dynamic layout means the right offer appears in the wrong position and loses much of its contextual advantage. Presentation and structure have to move together.
Layer 5 = Ancillary conversion: where the stack pays out
Ancillary revenue reached $148.4 billion globally in 2024. a record high, and a figure that now represents a structurally significant share of airline economics. The question for most airlines isn't whether ancillaries matter. It's whether they're capturing their fair share of that pool.
Conversion at this final layer depends on two things that the rest of the stack directly controls: the relevance of the offer (Layers 1 and 2) and the trust the passenger has in the purchase (Layer 4). But there's a third variable that unpacking ancillaries makes clear: flexibility and service-level transparency. An ancillary that can be cancelled or modified without friction converts at a higher rate than an identical product with rigid terms, because the perceived risk of purchase drops.
Critically, conversion at this layer feeds back into Layer 3. Every outcome — purchase, skip, abandon — is a signal the optimisation loop uses to sharpen the next recommendation. The stack is circular, not linear. Each completed journey through it makes the next one more accurate.
The stack is only as strong as its weakest layer
Most personalisation underperformance comes from investing heavily in one layer while neglecting the others. The patterns repeat across the industry:
- Strong segmentation data, weak decisioning leads to accurate audience, irrelevant offers
- Strong decisioning, no optimisation loop is a good starting point that gradually stales
- Live optimisation, static layout means the right offer buried in the wrong structure
- Everything working, poor ancillary SLAs means conversion undermined at the final step
The temptation is to start at Layer 4: dynamic banners and personalised images are visible, demonstrable, easy to screenshot for a board update. But visible personalisation without the foundation underneath it is expensive decoration. Research consistently shows that non-personalised, undifferentiated offers achieve less than 1% response rates. The gap between that baseline and what a properly-constructed stack produces is where the real commercial argument lives.
Build from the ground up. The payoff isn't visible until Layer 4, but the work that makes Layer 4 convert is done in Layers 1 through 3.
Interested in how Triplake supports personalisation across the full stack? Explore the platform here.
